I dropped my French tutor and built an LLM-based learning system

After months of paying $200 a month for private French lessons, the author realized the bottleneck wasn't teaching quality but two structural gaps: review timing and error specificity. This led to building a Claude-based knowledge tracker that uses the SM-2 spaced-repetition algorithm (the same one Anki uses) to schedule grammar reviews based on a personal forgetting curve, logging highly specific errors instead of vague categories.

Alongside it, the author built a browser-based voice app called Causons for real-time speaking practice, chaining three API calls: speech-to-text via Groq's Whisper-large-v3, chat via gpt-4o-mini, and text-to-speech via OpenAI's tts-1. The app steers conversations toward weak spots from the knowledge base and only corrects genuinely incorrect grammar, leaving colloquial expressions untouched.

Getting there took several cost and quality iterations: Azure's free tier turned out to push unwanted upgrades, Groq's Llama 3.3 model misflagged casual French as grammatically wrong, and both GPT-4o and Claude Sonnet were ruled out on cost grounds. The final stack runs on two API keys and costs just cents per session.

For engineers, this project is a concrete case study in designing personalized LLM-based learning tools, mixing providers strategically, and balancing cost against quality in a real, sustained use case.